All pull requests

Pull requests

• serve: opt-in single-engine preemption for short requests

  #2822·Zcg2021 opened Oct 16, 2026·

  ·
  1

• metal: --rope-scaling, --rope-scale, --rope-freq-base, --rope-freq-scale and --yarn-* reach the decode window's RoPE table

  #2819·oinani0721 opened Oct 16, 2026·

  ·
  1

  1

• bench: community report — 2x RTX 3060 12 GB, Unsloth UD-IQ4_XS, 131K context, agent-calibrated (engine 0.1.44.1)

  #2818·fenix79 opened Oct 16, 2026·3 tasks done·

  ·

• sycl: setup --vision cpu reads images on the Intel port (the CPU encoder, as on AMD)

  #2817·hotschmoe opened Oct 16, 2026·

  ·
  1

• strata-q36 on Volta: K-quant prompt experts on a register-fragment tensor-core kernel - prompt +22-25% on a V100, same tokens

  #2815·merbanan opened Oct 16, 2026Contributor·

  ·

• sycl: STRATA_GR_FUSED=1, the two-kernel BF16 hyper-connection read (opt-in)

  #2812·hotschmoe opened Oct 16, 2026·

  ·
  1

• sycl: STRATA_HC_Q8=1 reads the Q8_0 hyper-connection projections (it loaded them and never read them)

  #2810·hotschmoe opened Oct 16, 2026·

  ·
  1

• sycl: prompt attention over 8-query selection unions on XMX (opt-in STRATA_PROMPT_ATTN_UNION=1)

  #2808·hotschmoe opened Oct 16, 2026·

  ·
  1

• sycl: decode QSA selection and attention in fewer, shorter launches (opt-in, bitwise)

  #2806·hotschmoe opened Oct 16, 2026·

  ·
  1

• sycl: KREG small-batch mmvq kernels for Q4_K/Q5_K/Q6_K/IQ4_XS at 1-8 columns (opt-in STRATA_MMVQ_KREG=1)

  #2804·hotschmoe opened Oct 16, 2026·

  ·
  1

• sycl: cheaper IQ decodes in the grouped expert kernels (STRATA_IQ_DECODE_FAST=1, opt-in, bitwise)

  #2802·hotschmoe opened Oct 16, 2026·

  ·
  1

• HIP: verify-window routing for non-native routers (256 experts) on every HIP card

  #2799·pugolini opened Oct 16, 2026Contributor·

  ·

• bench: community report - 2x RTX 3090, IQ3_XXS, 262K context, parallel: 2

  #2796·saadharis opened Oct 16, 2026·

  ·

• bench: RTX 4090 community benchmarks for UD-Q4_K_XL, IQ3_XXS and IQ3_S (262K, one machine)

  #2795·hongxianzhi opened Oct 16, 2026·

  ·

• Community benchmark: RTX 3060 12 GB on an AVX-only Xeon E5-2630 v2 (no AVX2/FMA)

  #2793·Norhiro opened Oct 16, 2026·

  ·

• bench/results: RTX 5070 Ti (16 GiB) with 31.6 GiB of RAM — IQ2_XS at 128K in low-RAM mode

  #2791·Lithivm opened Oct 16, 2026·

  ·

• Use aria2c for model downloads

  #2790·hawkdriver992 opened Oct 16, 2026·

  ·
  1

• hip: route the asynchronous adaptive tier through adapt_copy_h2d (#884)

  #2788·iskamag opened Oct 16, 2026·

  ·

• docs: the 0.1.44.1 changes no page told: a crashed engine start names itself, and a setup run again keeps the installed size

  #2786·Buddychek opened Oct 16, 2026·

  ·

• HIP: the RX 6600 / 6650 XT (gfx1032) in setup's cards and CMake's unvalidated list

  #2784·chenxyz129 opened Oct 16, 2026·

  ·

• Start the vision encoder in parallel when it is on another GPU

  #2777·apatassini opened Oct 16, 2026·

  ·

• Expert arena load threads

  #2776·apatassini opened Oct 16, 2026·

  ·

• Per-client expert memory

  #2775·apatassini opened Oct 16, 2026·

  ·

• Short prompt chunks: router push + decode grouped kernel for resident experts

  #2774·apatassini opened Oct 16, 2026·

  ·

• Community benchmark: RX 7900 XTX (gfx1100), Ryzen 7 9800X3D - Qwen3.8-Flash-Next IQ2_XS, 131K

  #2773·neo20200528-sys opened Oct 16, 2026·

  ·